EP1646967B1 - Method for measuring the proximity of two contours and system for automatic target identification - Google Patents

Abstract

Method for measuring the proximity of a second contour (CM) relative to a first contour in which for each point (I k) of the first contour an association step with a point (M i) of the second contour, determined as the closest to it, is carried out. As a result each point on the second contour is coupled with a single or no point on the first contour. The invention also relates to a corresponding identification system.

Description

The present invention relates to the automatic identification of the targets present in an image. More specifically, this invention describes a discriminant method for comparing 2D contours. It applies mainly in the military field, to assist the pilot of a plane in a combat situation in his choice of shots. It is also of interest in any other field concerned with pattern recognition, in particular the field of surveillance and the medical field.

An automatic identification process must reliably determine how many targets are in the image, at what positions and what types they are.

A target is a 3D object that one seeks to identify. In the military field, these targets are typically tanks, land vehicles .... In the following, we speak indifferently target or object.

In this application, the term identification system, a system by which a target in an image is identified by its type: mark, name or number, or by its class: car, tank, car ...

The automatic identification of objects or targets is a complex algorithmic problem, partly because of the potential resemblances between two different targets from certain angles of view, and secondly because of the great variability of appearance of a target, due to geometric deformations, the position of certain elements, or the presence of certain equipment. For example, a vehicle may have open or closed doors, luggage on the roof ...

We try to identify automatically as reliably as possible targets in an image. The automatic identification process must thus have two essential qualities: to be robust, ie insensitive to variations in the appearance of a target that result in local disturbances on the object in the image; to be discriminating, that is to say to be able to discern between two targets close in appearance.

In the invention, one is more particularly interested in an automatic identification system of targets based on the comparison of contours. In such a system, the contours are initially extracted. present in the image to be analyzed, and then, in a second step, these contours are compared to those of a target reference database, containing data representing the 3D objects that one seeks to identify.

The extraction of the contours present in the image is done using a so-called segmentation technique. The result is an image called extracted outlines, corresponding to a binary image leaving only pixels of outlines, usually represented by white dots on a black background. In this image, only the outlines contain information. In the following, unless explicitly stated otherwise, the term "point" means a point carrying information, ie a point belonging to an outline in the model or in the image. Pixels that are not contour points do not carry information.

The extracted contour image is then compared to the contours obtained from a database representing the 3D objects that we are trying to identify. These outlines are called contour-models and are obtained, for each of the 3D objects, by projection according to a set of points of view making it possible to represent all the appearances of the object. Each 3D object in the database thus corresponds to a collection of model contours of this object.

In the invention, one is more particularly interested in a so-called correlative comparison method, which consists in comparing each model contour with the image of extracted contours for all the possible positions of this model contour in the image. For a given position, this comparison is performed by superimposing the model contour on the image, and consists in measuring "the difference" between the points of the model contour and those of the extracted contour image. Since each of the model contours is identified with respect to an origin, it is possible to recalculate the coordinates of each of its points in the coordinate system of the contour image, according to the pixel of the image on which this origin is centered. . Thus, each of the model contours is scanned over the entire contour image extracted.

When the extracted contour image has been scanned by the set of model contours, the process consists in selecting the most likely hypothesis or hypotheses.

By hypothesis, we mean a target, a position of this target in the image and a point of view under which this target is observed.

A method of estimating the difference between the model contour points and the extracted contour points is to count the number of points that these contours have in common.

This simple evaluation method based on the number of points in common with a model contour is, however, not very robust and not very discriminating. Not very robust because it is very sensitive to the variations of appearance of the target and not very discriminating because it takes into account with the same importance all the points of the outline.

Another more complex evaluation method uses a so-called Hausdorff measurement method. This method consists in identifying for each of the model contour points, the lowest distance from this point to the points of the image contour, and to deduce therefrom a degree of dissimilarity between the model contour and the image contour, on the basis of the average evaluated distances.

However, this process, even if it is more efficient than the previous one, is not robust enough nor discriminating, because it can take into account irrelevant distances that should be discarded. Indeed, the same model outline point can be seen as the closest to several different image outline points. This is particularly the case if the image contains parasitic points that do not correspond to a contour of a target to be identified, for example, points that correspond to internal contours of the target, or points that correspond to the target. environment of the target (vegetation, buildings ....). These parasitic points will disturb the measurement. Taking all these distances into account can thus lead to a false assumption.

An object of the invention is an automatic identification process that does not have these various disadvantages.

An automatic identification process according to the invention comprises a method for measuring the proximity of a model contour to an image contour based on a single matching step of each point of a model contour to zero or a single contour point picture.

This point-to-point matching method comprises a step of associating with each point of the image contour, the model contour point the closest. At this stage, two points of information are mapped to each point of the image contour: the coordinates of a model contour point determined as the closest and the distance between the two points thus associated.

Then, inversely, for each point of the model contour, we consider all the image contour points associated with it in the previous step and in this set, we determine the nearest image contour point, taking the lowest distance. We get a point-to-point univocal match. At the output, each model contour point is matched either to zero image contour point or to a single image contour point corresponding to a smallest distance.

By assigning a local proximity note to each model contour point, equal to zero if it is matched to zero image edge points, and if it is matched to an image edge point, equal to a value that is lower as the distance between the two paired points is large, it is possible to calculate an overall score, equal to the average of the local notes which expresses the probability of similarity of the model contour to the image contour.

The overall score resulting from this method is much more discriminating than the proximity measurement used in the automatic identification methods of the state of the art, especially with respect to false assumptions.

An automatic identification system according to the invention uses this method for each position of the model contour in the image, and for each model of a model collection.

The set of global scores obtained, corresponding to the different model contours and their different positions in the image, makes it possible to elaborate a certain number of hypotheses by retaining the best global proximity ratings.

The point-to-point matching process according to the invention makes it possible to improve the discrimination of the automatic identification system with respect to the false assumptions corresponding to cases where the contours in the image comprise interior points of contours. to say corresponding to the internal contours of a target, and external points of contours, ie corresponding to the environment of the target (vegetation, buildings ....).

According to another aspect of the invention, to improve discrimination between overlapping target hypotheses (ie at identical or near positions in the image, which is usually defined by contour points in common between the two hypotheses of model contours), the method of proximity measurement applies a local weighting at each point of a model contour. This weighting is representative of a quantity of information contained at this point and defined with respect to the other model contour. This weighting makes it possible to discriminate the silhouettes of the two targets on the basis of their local differences. More particularly, this weighting consists in applying the method of measuring proximity between the two model contours to be discriminated, in order to obtain, for each model contour, a weighting factor at each point which makes it possible to give more weight to the model contour points which contain the difference information with the other template outline. When the hypothesis collection contains more than two superimposable hypotheses, we apply this weighting process two by two, and we retain the best overall score obtained each time.

The automatic identification system according to the invention applies to each of the model contours of a collection, the process of measuring the proximity of this model contour to the image contour to be analyzed in order to evaluate the likelihood of this model, and between the model contours taken. two by two in a selection of overlapping hypotheses, to discriminate between two close model contours by locally weighting this probability relative to each of the two models.

Thus, as characterized, the invention relates to a method of measuring proximity of a second contour to a first contour, comprising for each point of the first contour, a step of association with a point of the second contour determined as the nearest , characterized in that it comprises a step of pairing each point of the second contour with one or zero points of the first contour, by determining the point of the first closest contour among the set of points of the first contour associated with said point of the second contour.

The invention also relates to a method for automatically identifying targets, which uses such a method for measuring the proximity of a model contour to an image contour.

According to one improvement, this identification method uses this method of measuring proximity of a model contour to another model contour, to allow discrimination between two overlapping hypotheses.

• la figure 1 représente une image de contours extraite d'une image d'entrée appliquée à un système d'identification automatique de contours;
• la figure 2 illustre l'étape d'association d'un point de l'image à un point de contour modèle selon un procédé de mesure de proximité d'un contour modèle au contour image à analyser selon l'invention;
• la figure 3 illustre l'étape d'appariement point à point selon un procédé de mesure de proximité d'un contour modèle au contour image à analyser selon l'invention ;
• les figures 4a et 4b illustrent un problème de détection d'hypothèses fausses ;
• les figures 5a et 5b illustrent les classes d'orientation associées aux points de contour image et modèle ;
• la figure 6 représente la courbe associée à un exemple de fonction d'attribution d'une note locale de proximité selon l'invention ;
• les figures 7a à 7d illustrent le principe de pondération de la note locale de pondération selon l'invention.

Other advantages and characteristics of the invention will appear more clearly on reading the description which follows, given by way of indication and not limitation of the invention and with reference to the appended drawings, in which:

• FIG. 1 represents a contour image extracted from an input image applied to an automatic identification system of contours;
• FIG. 2 illustrates the step of associating a point of the image with a model contour point according to a method for measuring the proximity of a model contour to the image contour to be analyzed according to the invention;
• FIG. 3 illustrates the step of point-to-point pairing according to a method for measuring the proximity of a model contour to the image contour to be analyzed according to the invention;
• Figures 4a and 4b illustrate a problem of detection of false assumptions;
• FIGS. 5a and 5b illustrate the orientation classes associated with the image and model contour points;
• FIG. 6 represents the curve associated with an exemplary function for assigning a local proximity note according to the invention;
• FIGS. 7a to 7d illustrate the weighting principle of the local weighting note according to the invention.

Figure 1 shows an image of outlines extracted from a data image, which may be from an infrared camera, a video system, or any other image source.

We want to determine in this image outlines extracted, how many targets it contains, which positions and which types, among a set of identified targets, that we have in the form of 3D objects in a database. For this, we build a set of 2D model contours corresponding to projections of each of the 3D objects, according to different angles of view, taking into account the information on the conditions under which targets are observed, for example, distance information between the target and the sensor, viewing angle ...

Consider a model outline positioned in any way in the extracted contour image, and noted CM. In the following image contour CI, the set of contour points of the extracted contour image is called. A proximity measurement method according to the invention is applied to measure the proximity of this contour CM to the image contour to be analyzed.

FIG. 1 illustrates an image of outlines extracted from an image obtained in any way: infrared image, active image. It contains image contour points, which correspond to the black pixels on this image, such as the points referenced I _a , I _b , I _c , I _d in FIG. 1. These contour points may be contour points of a target to be identified as the point I _a, the outer points to the contour of the target to be identified, such points I _b and I _c or points of an inner contour to the target to be identified, such as the point I _of .

The proximity measuring method according to the invention comprises a step of unambiguous pairing of each of the model contour points to zero or a single image contour point and a step of assigning a local proximity note to each point of the image. model outline, representing the proximity of this model outline point with the image outline.

• a)-une étape d'association à chaque point de contour image d'un point de contour modèle, sur le critère de la distance la plus faible ;
• b)-pour chaque point de contour modèle, la détermination de l'ensemble des points de contour image auxquels il a été associé à l'étape a)-, et la détermination du point image le plus proche dans cet ensemble, sur le critère de la distance la plus faible.

More specifically, the step of matching each of the points of the model contour comprises the following steps a) - and b) -:

• a) a step of associating with each point of the image contour of a model contour point, on the criterion of the weakest distance;
• b) for each model contour point, the determination of the set of image contour points to which it has been associated in step a), and the determination of the nearest image point in this set, on the criterion the smallest distance.

This method entails that, in step a), two information items are memorized for each image contour point: the coordinates of the associated model contour point and the corresponding distance between the two associated points, to perform step b) - d matching on the basis of these two pieces of information.

The distance considered is the Euclidean distance, of which a true measurement is measured or a discrete measurement according to the calculation methods used. In particular, the use of a chamfer method, making it possible in a known manner to accelerate the calculation time, uses a discrete measurement of the Euclidean distance.

Steps a) - and b) - are illustrated in Figures 2 and 3.

Step a) -. is illustrated in FIG. 2. The proximity of the image contour points CI to the model contour CM is evaluated to associate with each contour point a closest model contour point. Thus, as shown in FIG. 2, if a point of the image contour CI is taken, the evaluation of the nearest model contour point consists in finding the lowest distance d between this image contour point and an outline point. model. In the example shown in FIG. 2, this evaluation leads to associating the point M ₁ of model contour CM with the point I ₁ of the image contour C1. In this example, we also have the following associations: (I ₁ , M ₁ ) , (I ₂ , M ₁ ), (I ₃ , M ₁ ), (I ₄ , M ₂ ), (I ₅ , M ₂ ), (I ₆ , M ₃ ).

During this step, a same model contour point may be associated with different image contour points. In the example, the point M ₁ of the model contour CM has been associated with the image contour points I ₁ , I ₂ , I ₃ .

Step b) - is illustrated in Figure 3. It consists for each model contour point to select from the image contour points associated with it in the first step a), the nearest contour point image of the model outline point. In FIG. 3, the dotted lines represent the mapping of image contour points with model contour points according to the first step a). For each model contour point, there is thus 0, one or n associated contour points according to this step a). For example, for the image contour point M ₁₅ , there are 3 associated image contour points: l ₂₄ , l ₂₈ , and l ₂₉ .

Step b) consists in keeping only the nearest image point, when it exists, among the image contour points associated with the same model contour point and in evaluating the local proximity note of this model contour point. to the image contour on the basis of the matching (contour point model-contour point image) thus performed.

In the example of FIG. 3, the pairing point M _i (pattern) with point l _k (image) according to the invention is the following: (M ₁₀ , zero point image); (M ₁₁ , 1, ₂₀ ); (M ₁₂ , ₁₂ ); (M ₁₃ , ₂₂ ); (M ₁₅ , 1, ₂₄ ).

With a point-to-point pairing according to the invention, the image contour points I ₂₅ to I ₂₉ will therefore not be taken into account in the evaluation of the model's proximity.

The point-to-point matching step according to the invention provides for each point M _i of model contour M _i matched to a single image contour point I _k , a measurement of proximity of this point M _i to the image contour. This measure of proximity of the point M _i can be written as:

Dist (M _i ) = d (M _i , I _k ), where d (M _i , I _k ) is a true or approximate measure of the Euclidean distance between the paired points. It is expressed in number of pixels.

• si un point de contour modèle n'est atteint par aucun point de contour image, correspondant à un point de contour modèle très éloigné du contour image, on lui attribue la note zéro. Dans l'exemple de la figure 3, la note attribuée au point M₁₀ est zéro : N(M₁₀)=0.
• si un point de contour modèle est atteint par un point de contour image, unique, on lui affecte une note d'autant plus grande que les points sont proches. Par exemple, on pourrait avoir N(M₁₂)=0,7 ; N(M₁₅)=0,3.

The method further comprises a step of assigning a local proximity note to each of the points of the model contour as follows: the note takes a value between 0 and 1, all the greater the matched points are close (that the proximity measure of this point is weak). More precisely :

• if a model contour point is not reached by any image contour point, corresponding to a model contour point far away from the image contour, it is given the zero score. In the example of FIG. 3, the score assigned to the point M ₁₀ is zero: N (M ₁₀ ) = 0.
• if a model contour point is reached by an image boundary point, unique, it is assigned a note that is all the greater as the points are close. For example, we could have N (M ₁₂ ) = 0.7; N (M ₁₅ ) = 0.3.

The last step of the process is then to determine the overall score for the model, by averaging the local scores of all the points of the model outline.

According to this evaluation principle, the model contour is evaluated as being all the closer to an image contour that the overall rating assigned to it is high.

It has been possible to show that such an automatic target identification process according to the invention makes it possible to avoid detection errors of the type illustrated in FIGS. 4a and 4b. In these figures, we superimposed on an image 1 comprising a target C, a first model MOD ₁ (Figure 4a) and a second model MOD ₂ (Figure 4b). The first model MOD ₁ corresponds in the example to the target to detect on which it is perfectly positioned. It leads to a hypothesis retained. The second model corresponds to another type of target. But with a method according to the state of the art, the hypothesis will be retained, because of the presence of contour points not belonging to the contour of the target, but actually belonging to the bottom, or to contour points internal.

According to one embodiment of the invention, the evaluation of the local proximity note of each model contour point is a function of the distance d between this point and the paired image contour point according to the invention.

Preferably, and as shown diagrammatically in FIGS. 5a and 5b, the evaluation of the proximity of two points comprises taking into account the orientation class at the points I ₂₀ and M ₃₀ of the pair P considered. This class is typically defined by the orientation of the tangent of the contour at the point considered: in FIG. 5a, the tangent t _{I is represented} at the image contour point I ₂₀ of the image contour CI and the tangent t _M at the contour point model M ₃₀ of the contour model CM. We define n orientation classes, with n integer: the orientation class 0 corresponds to a horizontal orientation of the tangent; the orientation class n-1 corresponds to a vertical orientation of the tangent and each of the intermediate orientation classes corresponds to an orientation of the determined tangent between 0 and π rad. These classes are shown in Figure 5b with n = 8. In this example, the ₂₀ point I belongs to the orientation class 6 and the point M ₃₀ is part of the orientation class 5.

In general, if the tangents t _I and t _M coincide, that is, if the paired points belong to the same orientation class, then ΔORI = 0. If the paired points are in orthogonal classes, ΔORI = n-1. More generally, we have AORI = | class (Ik) -class # (Mi) | (in number of pixels).

In the example shown in FIG. 5a , ΔORI = 6-5 = 1

The corrected proximity measurement to the image contour of the model contour point M _i matched to the image contour point l _k can thus be written as follows: $$\mathrm{dist}\left({\mathrm{M}}_{\mathrm{i}}\right)=\mathrm{d}\left({\mathrm{M}}_{\mathrm{i}},,{\mathrm{I}}_{\mathrm{k}}\right)+\frac{1}{4}\mathrm{\Delta ORI}\mathrm{.}$$

In practice, with n = 8 classes, we obtain a good compromise in terms of false detections and computing time.

In this improvement, proximity measurement is a continuous function of position and orientation. This limits the weight of the orientation, which can be erroneously estimated.

In an alternative taking into account of the orientation class, the orientation class is taken into account in the step of associating the point-to-point matching process, by not allowing the association (and therefore the pairing) only between points of the same class. In this case, the proximity measure Dist (M _i ) is equal to the distance between the paired points M _i and l _k .

The attribution of the local proximity note N (M _i ) of a model contour point M _i as a function of the proximity measurement according to the invention must contribute to the robustness of the identification method.

This local note translates a probability of similarity between the model contour and the image contour: it takes a value on the interval [0,1]. When it is zero, it translates that the model contour point does not "match" with the image outline; when it is equal to 1, it reflects a high probability that the model contour corresponds to the image contour.

Thus, all the model contour points that could not be matched with an image contour point according to the method of the invention must have a zero contribution, that is to say a zero note, reflecting their large distance from the image contour .

• la note doit prendre une valeur égale à 1, lorsque la mesure de proximité Dist(M_i) est nulle ;
• la note doit prendre une valeur voisine de 1 lorsque la mesure de proximité Dist(M_i) est comprise entre 0 et 1.
• la note doit décroître très rapidement vers 0 dès que la mesure de proximité Dist(M_i) devient supérieure à 1.
• la courbe d'attribution de la note N(M_i) possède un point d'inflexion, de préférence pour une mesure de proximité Dist(M_i) voisine de 2 pixels.
• la note doit prendre une valeur quasi-nulle dès que la mesure de proximité Dist(M_i) devient supérieure à 3 pixels.

For model contour points that are matched to a single, contoured point, the score assignment function preferably follows the following criteria:

• the note must take a value equal to 1, when the proximity measure Dist (M _i ) is zero;
• the note must take a value close to 1 when the proximity measure Dist (M _i ) is between 0 and 1.
• the note must decrease very rapidly to 0 as soon as the proximity measure Dist (M _i ) becomes greater than 1.
• the allocation curve of the note N (M _i ) has a point of inflection, preferably for a proximity measurement Dist (M _i ) close to 2 pixels.
• the note must take a quasi-zero value as soon as the proximity measure Dist (M _i ) becomes greater than 3 pixels.

The function N (M _i ) for assigning the note to a model contour point M _i paired according to the invention at the image contour point I _k will have for example the shape shown in FIG. 6, which corresponds to the following function : $$\mathrm{NOT}\left({\mathrm{M}}_{\mathrm{i}}\right)\mathrm{=}\left(\mathrm{0}\mathrm{,}\mathrm{5}-\mathrm{<figure-callout\; id="4"\; label="arctan"\; filenames="imgf0003.png"\; state="\{\{state\}\}">arctan</figure-callout>}\frac{\mathrm{4}\left(\mathit{dist}\left({\mathit{M}}_{\mathit{i}}\right)\mathrm{-}\mathrm{2}\right)}{\mathit{\pi }}\right)\frac{\mathrm{1}}{\mathrm{0}\mathrm{,}\mathrm{9604}}\mathrm{.}$$

A practical implementation of a proximity measurement method according to the invention can use so-called chamfer calculation methods. These chamfer methods are very efficient in terms of computation time and are widely used in many areas of image processing including that of shape recognition.

A classical chamfer method allows to map to a contour, a map with two inputs x and y corresponding to the coordinates of a given point, and an output, which is the lowest distance from this point (x, y) to the outline. In other words, the lowest distance from the point (x, y) to the contour mapped by contour lines is evaluated. This known method of chamfer is generally used to apply the Hausdorff measurement method. In this case, the chamfer method is applied to the image contour, making it possible to determine for each point (x, y) of model contour, the smallest distance to the image contour.

In the method according to the invention, the chamfer method must be applied differently.

First, in the first association step of the method according to the invention, the aim is to measure the smallest distance from an image contour point to the model contour. This involves applying the chamfering method no longer to the image contour, but to each of the model contours.

However, the calculation of the chamfer map of a model contour is independent of the extracted contour image to be analyzed. These calculations can be done once and for all and stored, to be used when the time comes, in real time, for the analysis of a given contour image.

Then, to allow point-to-point matching according to the invention, the chamfer map of the model contour must output a first information which is the distance between the two associated points, and a second information which is the identification of the model contour point associated with this distance. This second piece of information is necessary because it will enable the pairing stage to determine all the image contour points associated with the same model contour point, and to automatically deduce the proximity measure. by the first associated information.

Thus, a rapid calculation method according to the invention comprises the calculation of a chamfer map for each model contour, said map giving as a function of the two inputs x and y corresponding to the coordinates of an image contour point, an information S ₀ (x, y) identifying the model contour point reached by the lowest distance measurement, and information S ₁ (x, y) corresponding to the value of this measurement.

Then, we apply the steps of point-to-point matching, and assigning a local note to each model contour point, depending on the proximity measure Dist (M _i ) for the paired points.

This method does not make it possible to correct the distance measure Dist (M _i ) of a model contour point M _i as a function of the orientation class of this point and of the paired image contour point I _k .

It is then expected to calculate a chamfer map by orientation class of the model contour. We therefore have n chamfer maps by model contour. We have seen that, preferably, n = 8.

The step of associating with each contour point of a model contour point then comprises, for each image contour point, the preliminary determination of the orientation class of this point, and the selection of the chamfer card. of the model contour in the corresponding orientation class.

Finally, the calculation of the overall score η consists in averaging all the local scores, ie, if the model contour comprises the points M _{i = 1 to l} of model contour, $\eta =\frac{1}{l}{\displaystyle \underset{i=1}{\overset{l}{\Sigma }}}\mathrm{NOT}\left(M_i\right)\mathrm{.}$

The method according to the invention is applied to all model contours by scanning each time on the entire image.

An overall score is obtained for each model contour (model contour to be understood as the model contour in a given position), which is a probability measure of similarity of this model contour to the image contour.

For example, we obtain the overall score η ₁ for the model contour CM ₁ ; η ₂ for the model contour CM ₂ , ...

Preferably, a selection of hypotheses is then established. By hypothesis is meant a model contour (that is to say a target, from a certain point of view) in a determined position in the image.

The selection is typically obtained by retaining the most probable assumptions corresponding to an overall score obtained above a decision threshold. This threshold is preferably set at 0.6.

The implementation of such an automatic target identification process using a proximity measurement method according to the invention makes it possible to reduce the number of false alarms and to better discriminate between the various hypotheses. In other words, we have fewer hypotheses at the output.

The table above shows, for comparison, for different images containing only one target to be identified, the number of hypotheses retained on the Hausdorff measurement criterion (hypothesis retained if Hausdorff measurement <2pixels) and on the criterion of the overall proximity score (η> 0.6) according to the invention. We see that the selection criterion based on the proximity note according to the invention gives much better results in terms of rejection of false assumptions. Image 1 Image 2 Image 3 Image 4 Image 5 Image 6 Image 7 Image 8 Hausdorff 4 5 3 2 3 7 2 8 Overall rating η 3 2 0 2 3 2 0 4

On the other hand, it does not make it possible to improve in a really decisive way the discrimination between two targets of close silhouettes. This results in the presence of superimposable hypotheses in the hypothesis selection obtained. The notion of overlapping hypotheses is a a concept well known to those skilled in the art. It shows that the model contours of these hypotheses have contour points in common.

Another aspect of the invention makes it possible to improve this last point.

The problem more particularly considered here is due to the fact that, according to certain angles of view, for certain orientations, two targets may have relatively similar shapes, similar to the meaning of the overall score η assigned according to the invention.

Nevertheless, one can notice in practice the presence of localized differences. For military vehicles for example, it may be the presence of tracks or wheels; a significantly different length; a rounded or angular shape ...

Some parts of a model outline are therefore more informative than others with respect to another model outline.

FIG. 7d thus shows an IC image contour corresponding to an image of extracted contours. Two models CM ₁ and CM ₂ shown respectively in Figure 7a and Figure 7b, are found close to the meaning of the invention of this image contour.

The basic idea of the improvement according to the invention consists in considering the two hypotheses which are superimposed, and in weighting the local note of each point of a model contour, established in the measure of proximity of this model contour to the image contour. , by a quantity of information representing the local difference at this point, with the other model contour.

According to the invention, the overall score η ₁ associated with the model contour CM ₁ which measures the probability of similarity of this model contour CM ₁ to the image contour CI is obtained by weighting each of the local notes. More precisely, the local proximity note N (M1 _i ) of each point M1 _i of the model contour CM1 is weighted by a factor representative at this point of the quantity of discriminant information that it contains relative to the other contour. CM model ₂ . This amount of information contained in a point M1 _i of the model contour CM ₁ must be all the more high since this point is far from the other model contour CM ₂ : it is the very definition of the proximity measurement in this Dist point (M1 _i ) according to the method of the invention.

où M2_j est un point du contour modèle CM2_j apparié au point M1_i selon le procédé de mesure de proximité de l'invention. En un point donné du contour modèle CM₁, plus la distance au point apparié est grande, plus la quantité d'information en ce point est importante. C'est ce que représente schématiquement la figure 7c. Au point M1_a, la quantité d'information X(M1_a) est grande, correspondant à la distance d_a sur la figure.The amount of information of each of the points M1 _i of the first contour CM ₁ relative to the contour CM ₂ is thus defined as follows: $$\mathrm{X}\left({\mathrm{M}\mathrm{1}}_{\mathrm{i}}\right)\mathrm{=}\mathrm{dist}\mathrm{\left(}{\mathrm{M}\mathrm{1}}_{\mathrm{i}}\mathrm{\right)}\mathrm{=}\mathrm{d}\mathrm{(}{\mathrm{M}\mathrm{1}}_{\mathrm{i}}\mathrm{,}{\mathrm{M}\mathrm{2}}_{\mathrm{j}}\mathrm{)}\mathrm{.}$$

where M2 _j is a point of the model contour CM2 _j paired with the point M1 _i according to the proximity measuring method of the invention. At a given point of the model contour CM ₁ , the greater the distance to the matched point, the greater the amount of information at this point. This is schematically represented in Figure 7c. At the point _M1, the amount of information X _(M1) is large, corresponding to the distance _a in Fig.

At the point M1 _b, the amount of information X (M1 _b) is zero, since this point, the two contours merge.

The methods for calculating the chamfer and for taking into account the orientation of the points in the distance measurement described above apply in the same way to this calculation of the amount of information.

The weighting method according to the invention then consists, in the step of calculating the overall score η ₁ of the contour CM ₁ to the image contour, to weight the local proximity note of each point M 1 _i of the model contour CM ₁ by the quantity of information X (M1 _i ) associated, that is: $${\mathrm{<figure-callout\; id="1"\; label="\eta "\; filenames="imgf0003.png"\; state="\{\{state\}\}">\eta </figure-callout>}}_1=\frac{1}{m}{\displaystyle \underset{i=1}{\overset{m}{\Sigma }}}\mathrm{NOT}\left({M1}_i\right)\mathrm{.}X\left({M1}_i\right)\mathrm{.}$$

The weighting method is applied to the points of the second contour CM ₂ , inverting the role of the first and second contours, that is to say using the method of measuring the proximity of the second contour CM ₂ to the first contour CM ₁ : we obtain the amount of information X (M2 _j ) of each point M2 _j of the second contour CM ₂ relative to the first contour CM ₁ . The local proximity note N (M2 _j ) of each point M2 _{j is} weighted by the associated quantity of information X (M2 _j ). The overall score η ₂ is obtained by averaging the weighted local proximity score of each of the points of the model contour CM ₂ , ie $${\eta }_{}$$$${}_2=\frac{1}{l}{\displaystyle \underset{j=1}{\overset{l}{\Sigma }}}\mathrm{NOT}\left({M2}_j\right)\mathrm{.}X\left({M2}_j\right)\mathrm{.}$$

Thus, more weight is given to the parts of the model outline that have the most information compared to others.

In other words, it amounts to discriminating between the two hypotheses on the basis of the contour points that contain the most information compared to the other.

This notion of quantity of information is thus defined with respect to a given couple of model contours.

When more than two hypotheses overlap, the discrimination process is applied two by two.

Thus, the invention describes a method for measuring the proximity of a second contour to a first contour, according to which each point M _i of the second contour is matched with one or zero points of the first contour, giving a distance measurement Dist (M _i ) at this point.

An automatic target identification method according to the invention applies this proximity measurement process to determine the measure of proximity of each point of a model contour, applied as a second contour, to an image contour, applied as the first contour. He deduces for each point of the model contour, a local note of proximity and for the model contour, a global note, giving a measure of likelihood of similarity to the image contour.

The automatic identification method thus determines the overall score associated with each of the model contours of a collection (with as many different model contours as different 3D models and points of view considered for each 3D model).

According to another aspect of the invention, it applies a criterion of selection of hypotheses, retaining as probable hypothesis, each of the model contours whose overall score is greater than the threshold.

According to one variant, the model contours of the collection correspond to a selection of hypotheses, resulting from another process, for example, resulting from a Hausdorff measurement.

According to another aspect of the invention, the automatic identification method then applies the weighting method to each pair of hypotheses which are superimposed among the assumptions retained, in order to obtain for the model contour associated with each hypothesis a weighted overall score. according to the invention. For this, he uses the method of measuring proximity, applying it a first time, to measure the amount of information associated with each point of the contour of the first hypothesis, applied as a second contour, relative to the outline of the second hypothesis. applied as the first contour, and calculate the overall rating associated by averaging the weighted local scores. It applies the proximity measurement method a second time to measure the amount of information associated with each point of the contour of the second hypothesis applied as the second contour, relative to the contour of the first hypothesis applied as the first contour, and to calculate the overall score. associated by averaging the weighted local scores. Then the identification system selects the best hypothesis. If the hypotheses that are superimposed are greater than two, the automatic identification system applies this weighting two by two, to retain each time the best hypothesis.

It has been possible to test on a workstation the performance of an automatic identification system for targets using such an identification method according to the invention, on a base containing 200 images to be analyzed and 9 3D targets corresponding to terrestrial vehicles. It has thus been possible to demonstrate a significant improvement in the performance of the identification, with a good identification rate of 80%, compared to 50% obtained with state-of-the-art methods.

It should be noted that the application of an automatic identification system according to the invention to a first selection of hypotheses obtained from another automatic identification process, such as a process using the Hausdorff measurement, does not change. identification performance, but advantageously saves computing time.

The invention which has just been described makes it possible to appreciably improve the robustness and the discrimination of an automatic identification system which implements it. It applies to the military field, but more generally, to any domain using pattern recognition as compared to a series of models.

Claims (19)

1. Automated method of measurement of proximity of a second contour (CM) extracted from an image to a first contour (CI), comprising for each point (I_k) of the first contour, a step of association with a point (M_i) of the second contour determined as the closest, characterized in that it comprises a step of pairing each point of the second contour with one or zero points of the first contour, by determining the point of the first contour which is closest from among the set of points of the first contour that are associated with said point of the second contour.
2. Method according to Claim 1, characterized in that the determination of a point that is closest to a given point is based on a true or discrete measure of the Euclidean distance between the two points.
3. Method according to Claim 2, characterized in that it comprises a step of allocating a measure of proximity Dist (M_i) of each point M_i of the second contour (CM) to the first contour (CI), based on the measurement of the distance from this point to the point of the first contour with which it is paired.
4. Method according to Claim 3, characterized in that said distance measure is a measure corrected as a function of the difference of class of orientation of the points of the pair considered.
5. Method according to any one of Claims 1 to 3, characterized in that in the step of associating zero or one points of the second contour with each point of the first contour, the point that is closest from among the points of the second contour which have the same class of orientation as said point of the first contour is associated.
6. Method according to any one of Claims 1, 2, 3, or 5, characterized in that the associating step uses a chamfer map of the second contour via which, at each point of the first contour with coordinates x and y applied as input, said map provides as output an identification (S₀(x,y)) of the point of the associated second contour and a measure (S₁(x,y)) of the proximity between the two points thus associated.
7. Method according to Claim 6 in combination with claim 5, characterized in that with the second contour is associated a chamfer map per class of orientation, and in that for each point of the first contour, the associating step comprises a step of determining the class of the point of the first contour, so as to apply the coordinates (x,y) of this point as inputs to the chamfer map corresponding to said orientation class.
8. Method according to any one of Claims 4 to 7, characterized in that it uses eight orientation classes.
9. Method of automatic identification of targets in an image of extracted contours (CI) which is defined by an image contour, characterized in that it applies a method of measurement of proximity according to any one of Claims 1 to 8, by applying as second contour, a template contour (CM) and as first contour, said image contour (CI), so as to obtain the measure of proximity Dist(M_i) of each point of said template contour to said image contour.
10. Method of identification according to Claim 9, characterized in that it comprises the allocation of a local score of proximity N(M_i) to each point M_i of the template contour as a function of the measure of proximity Dist(M_i) of this point, according to the following criteria:

    - N(M_i) has a value lying between 0 and 1.

    - N(M_i) =0, when said point is paired with zero points of the first contour;

    - N(M_i) = 1, when the proximity measure is equal to zero;

    - N(M_i) has a value of about 1 when the proximity measure lies between 0 and 1 pixels.

    - N(M_i) decreases very rapidly to 0 as soon as the proximity measure becomes greater than 1 pixel.

    - N(M_i) decreases according to a curve having a point of inflexion, in the neighborhood of a proximity measure of about 2 pixels.

    - N(M_i) has a quasi-zero value as soon as the proximity measure becomes greater than 3 pixels.
11. Method of identification according to the preceding claim, characterized in that the function for allocating the score of proximity to the point M_i may be written: $$\mathrm{N}\left({\mathrm{M}}_{\mathrm{i}}\right)\mathrm{=}\left(\mathrm{0.5}\mathrm{-}\arctan \frac{\mathrm{4}\left(\mathrm{Dist}\left({\mathrm{M}}_{\mathrm{i}}\right)\mathrm{-}\mathrm{2}\right)}{\mathrm{\pi }}\right)\frac{\mathrm{1}}{\mathrm{0.9604}}\mathrm{.}$$

    
12. Method of identification according to Claim 10 or 11, characterized in that it comprises a step of measuring a global score η equal to the mean of the proximity scores relative to the number of points of the template contour (CM).
13. Method of identification according to any one of Claims 9 to 12, characterized in that it is applied successively to each of the template contours of a collection of template contours.
14. Method of identification according to Claim 13, characterized in that said collection is obtained from another method of identification of targets, such a method using a Hausdorff distance measure.
15. Method of identification according to Claim 13 or 14, characterized in that it comprises a step of selecting hypotheses by comparison with a threshold of each of the global scores η allocated to each of the template contours.
16. Method of identification according to Claim 15, characterized in that said threshold is fixed at 0.6.
17. Method of identification according to one of Claims 15 or 16, characterized in that it comprises a step of discriminating between hypotheses of template contours which are superimposed, comprising for each pair of a first contour hypothesis (CM₁) · and of a second contour hypothesis (CM₂) which are superimposed, a step of weighting the global score allocated to each of the template contours, said weighting step comprising the application of the method of measurement of proximity according to any one of Claims 3 to 8:

    - a by applying as second contour, the contour of said first hypothesis and as first contour, the contour of said second hypothesis, said proximity measure Dist(M1_i) obtained for each point (M1_i) of contour (CM₁) of the first hypothesis being applied as weighting factor X(M1_i) for the local score of proximity (N(M1_i)) of this point to the image contour (CI), and by deducing the global score (η₁) associated with the first contour hypothesis representing its proximity to the image contour by calculating the mean of said weighted local scores,

    - b by applying as second contour, the contour of said second hypothesis and as first contour, the contour of said first hypothesis, said proximity measure Dist(M2_j) obtained for each point (M2_j) of contour (CM₂) of the first hypothesis being applied as weighting factor X(M2_j) for the local score of proximity (N(M2_j)) of this point to the image contour (CI), and by deducing the global score (η₂) associated with the first contour hypothesis representing its proximity to the image contour by calculating the mean of said weighted local scores.
18. Method of identification according to Claim 17, characterized in that it adopts as best hypothesis of template contour, from among a plurality of hypotheses which are superimposed, that with which the best global score is associated.
19. System for the automatic identification of an object in an image of contours, comprising a database containing templates of determined objects to be recognized, and means of calculation, said means of calculation being configured so as to compare contours, characterized in that said means of calculation are configured so as to perform the steps of a method of identification, according to any one of Claims 9 to 18.

Images